Meteorological satellite image noise classification detection method and system based on visual large model

Abstract

The present invention discloses a method and system for detecting noise levels of meteorological satellite images based on a large visual model, wherein the method comprises: obtaining a meteorological satellite image for preprocessing, and cropping the image according to a preset method to generate a plurality of meteorological satellite image blocks. Based on a preset recognition model, noise type recognition is performed on each of the meteorological satellite image blocks to generate a recognition result. Based on the recognition result, detection is performed using a preset detection model to generate a noise detection result. Thus, the method for detecting noise levels of meteorological satellite images based on a large visual model of the present invention improves the accuracy of noise detection by hierarchically identifying noise types and applying the corresponding large visual model for noise detection.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of satellite image technology, and in particular to a method and system for noise classification detection of meteorological satellite images based on a large visual model. Background Art

[0002] With the development of earth observation technology, meteorological satellite image data provides important information support for weather forecasting, disaster prediction and other fields. However, due to the complex and changeable space environment, meteorological satellite images are susceptible to electromagnetic interference during the acquisition and transmission process, resulting in noise such as noise points or noise lines in the image, which seriously affects the quality and availability of the image. Existing noise detection methods for meteorological satellite images usually use traditional shallow methods, which cannot effectively learn the deep semantic features of noise, resulting in low noise detection accuracy. Therefore, how to effectively identify and detect noise in meteorological satellite images has become an urgent problem to be solved. Summary of the invention

[0003] The embodiment of the present invention provides a method and system for hierarchical noise detection of meteorological satellite images based on a visual large model, which improves the accuracy of noise detection by identifying noise types in a hierarchical manner and applying corresponding visual large models for noise detection.

[0004] To achieve the above objectives, in a first aspect, the present invention provides a method for noise classification detection of meteorological satellite images based on a visual large model, comprising:

[0005] Step S100, obtaining a meteorological satellite image for preprocessing, and cropping it according to a preset method to generate a plurality of meteorological satellite image blocks;

[0006] Step S200, performing noise type identification on each of the meteorological satellite image blocks based on a preset identification model to generate an identification result;

[0007] Step S300: Based on the recognition result, a preset detection model is used to perform detection to generate a noise detection result.

[0008] In one embodiment of the present invention, step S100 includes:

[0009] Step S101, obtaining a large-scale meteorological satellite image;

[0010] Step S102, performing preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image;

[0011] Step S103 , based on the processed large-scale meteorological satellite image, cropping is performed according to a preset method to generate the plurality of meteorological satellite image blocks.

[0012] In one embodiment of the present invention, step S200 includes:

[0013] Step S201, obtaining a preset recognition model, wherein the preset recognition model includes a noise point image recognition model and a noise line image recognition model;

[0014] Step S202: Based on the noise point image recognition model and the noise line image recognition model, noise type recognition is performed on each meteorological satellite image block to generate a noise point recognition result, a noise line recognition result and a noise-free result.

[0015] In one embodiment of the present invention, step S300 includes:

[0016] Step S301, obtaining a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model;

[0017] Step S302, based on the noise recognition result, use the image noise detection model to perform detection to generate a noise detection result;

[0018] Step S303, based on the noise line recognition result, using the image noise line detection model to perform detection to generate a noise line detection result;

[0019] Step S304: generating the noise detection result based on the noise point detection result and the noise line detection result.

[0020] In a second aspect, the present invention provides a meteorological satellite image noise classification detection system based on a visual large model, comprising: a first generation module, a second generation module and a third generation module. The first generation module is used to obtain a meteorological satellite image for preprocessing, and to crop it according to a preset method to generate a plurality of meteorological satellite image blocks. The second generation module is used to identify the noise type of each meteorological satellite image block based on a preset recognition model to generate a recognition result. The third generation module is used to perform detection based on the recognition result using a preset detection model to generate a noise detection result.

[0021] In one embodiment of the present invention, the first generation module includes: a first acquisition unit, a first generation unit, and a second generation unit. The first acquisition unit is used to acquire a large-scale meteorological satellite image. The first generation unit is used to perform preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image. The second generation unit is used to perform cropping based on the processed large-scale meteorological satellite image according to a preset method to generate the multiple meteorological satellite image blocks.

[0022] In one embodiment of the present invention, the second generation module includes: a second acquisition unit and a third generation unit. The second acquisition unit is used to acquire a preset recognition model, wherein the preset recognition model includes a noise image recognition model and a noise line image recognition model. The third generation unit is used to identify the noise type of each meteorological satellite image block based on the noise image recognition model and the noise line image recognition model, and generate a noise recognition result, a noise line recognition result and a noise-free result.

[0023] In one embodiment of the present invention, the third generation module includes: a third acquisition unit, a fourth generation unit, a fifth generation unit and a sixth generation unit. The third acquisition unit is used to acquire a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model. The fourth generation unit is used to perform detection based on the noise point recognition result using the image noise point detection model to generate a noise point detection result. The fifth generation unit is used to perform detection based on the noise line recognition result using the image noise line detection model to generate a noise line detection result. The sixth generation unit is used to generate the noise detection result based on the noise point detection result and the noise line detection result.

[0024] In a third aspect, the present invention provides an electronic device, comprising:

[0025] at least one processor; and

[0026] a memory communicatively coupled to the at least one processor;

[0027] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the meteorological satellite image noise classification detection method based on the visual large model as described above.

[0028] In a fourth aspect, the present invention provides a computer-readable storage medium comprising a computer program and instructions. When the computer program or the instructions are executed on a computer, the computer executes the meteorological satellite image noise grading detection method based on a visual large model as described above.

[0029] Compared with the prior art, the method and system for detecting noise levels in meteorological satellite images based on a large visual model according to the present invention have the following beneficial effects:

[0030] 1. Through the noise point or noise line image recognition model based on the Transformer network, it is possible to accurately identify whether the meteorological satellite image contains noise points or noise lines, providing a basis for subsequent noise detection;

[0031] 2. Using the meteorological satellite image noise point or image noise line detection model based on the visual big model can effectively learn the deep semantic features of noise and improve the accuracy of noise detection;

[0032] 3. Through hierarchical identification and detection, the noise type is first identified and then the corresponding detection model is applied, which improves the efficiency and accuracy of noise detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of a method for noise classification detection of meteorological satellite images based on a large visual model in Embodiment 1 of the present invention;

[0034] Figure 2 It is a structural schematic diagram of a meteorological satellite image noise classification detection system based on a visual large model in Embodiment 2 of the present invention;

[0035] Figure 3 It is a schematic diagram of the structure of an electronic device in Embodiment 3 of the present invention;

[0036] Figure 4 It is a logic flow diagram of a noise classification detection method for meteorological satellite images based on a visual large model in a specific embodiment of the present invention;

[0037] Figure 5 It is a structural schematic diagram of an image noise point detection model and an image noise line detection model based on a large visual model in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention, rather than to limit the embodiments of the present invention. It is also necessary to explain that, for ease of description, only parts related to the embodiments of the present invention are shown in the accompanying drawings, rather than all structures.

[0039] To facilitate understanding, the main implementation concepts of the embodiments of the present invention are first briefly described.

[0040] With the development of earth observation technology, more and more meteorological satellites are being deployed, such as the Fengyun-3 meteorological satellite. These satellites are equipped with various sensors to collect information about the earth's atmosphere and provide comprehensive, timely and dynamic cloud images. These data are vital to the meteorological department and provide a rich data basis for disaster prevention and mitigation. Effective use of this information is essential for tasks such as weather forecasting, disaster prediction, and supporting agricultural production. However, the space environment is susceptible to electromagnetic interference, which makes satellite sensors susceptible to interference, resulting in the collected images containing various types of noise, significantly reducing their usability. How to detect these noisy images from massive meteorological satellite data is of great help and significance for improving the accuracy of meteorological satellite images in downstream tasks (such as crop yield assessment).

[0041] At present, the existing meteorological satellite noise images are mainly divided into two types: images with noise points and images with noise lines. Noise point images are similar to salt and pepper noise, which are usually caused by factors such as image sensor damage, transmission channel abnormalities, and abnormal decoding processes. Noise line images usually appear as black or white rectangular strips, which are usually caused by interference or failures in the imaging process. The current meteorological satellite noise detection methods usually determine whether there is noise in the meteorological satellite image before detection. Then, different noise detection methods are designed for different categories of noise images. These noise recognition and detection methods are mostly traditional shallow methods, which cannot effectively learn the deep semantic features of noise, and thus cannot accurately identify and detect different types of noise information.

[0042] Noise detection methods for meteorological satellite images are mainly divided into two types: noise points and noise lines. In terms of noise point detection, shallow strategies are often used to learn noise features. For example, the selective median filter (PSM) method compares the absolute difference between the original pixel grayscale value and the median filter value, and uses an empirical threshold to determine whether a pixel has noise. The judgment is made only when the difference exceeds the threshold. The three-state median filter (TSM) calculates the output of the standard median filter and the center weighted median filter (CWM) for each pixel, and then determines whether the absolute difference between these outputs and the pixel value is greater than the threshold T to determine whether the pixel is noise. The adaptive center weighted median filter (ACWM) calculates the difference between the value after weighted median filtering and the original pixel grayscale value, and uses multiple empirical ordered thresholds to iteratively identify potential pixel noise with different possibilities. The relative absolute difference (ROAD) algorithm calculates the absolute difference between the current pixel and the pixel value in its window, sorts the differences, and adds the four minimum values ​​to obtain ROAD, which effectively distinguishes bad pixels from edges. The logarithmic relative difference ranking (ROLD) algorithm applies logarithmic operations based on ROAD, and uses the mean and mean square error to determine whether the current point is a bad pixel. The main feature of a noise line is that it appears as a straight line and is relatively different from the surrounding pixels. Noise line detection methods mainly include threshold-based methods, Canny-based line detection, and Hough transform-based methods. The threshold-based method relies on the significant difference characteristics between the bad line and the surrounding pixels, calculates the pixel difference between the lines, and uses a specified threshold to determine whether the pixel difference between the two lines is greater than the threshold, thereby classifying it as a bad line. Canny-based noise line detection first applies the Canny operator to obtain the edge of the image, then calculates the number of edge pixels in the vertical direction to determine whether it is within the specified threshold range, and considers horizontal edges with a distance less than or equal to 5 as bad lines. Hough transform-based line detection modifies the horizontal edge determination method in Canny-based line detection, using a more accurate Hough transform for this purpose.

[0043] Most of the existing noise detection methods for meteorological satellite images use traditional shallow methods, which cannot effectively learn the complex semantic features of noise and thus cannot accurately identify and detect different types of noise information.

[0044] The inventors have discovered the defects of the prior art as mentioned above, and have provided a method and system for noise grading detection of meteorological satellite images based on a visual large model. First, a hierarchical recognition model of noise images is trained using a Transformer network to effectively identify whether the meteorological satellite image contains noise points or noise lines. Then, for images containing noise points or noise lines, the present invention uses the corresponding large visual models to detect the noise points and noise line areas in the image. By first identifying the noise in a graded manner, and then independently using the large visual model to detect the noise data, the present invention can effectively improve the accuracy of noise detection and effectively improve the shortcomings of existing noise detection methods.

[0045] Embodiment 1

[0046] Figure 1 is a flow chart of a method for noise classification detection of meteorological satellite images based on a visual large model in Embodiment 1 of the present invention, such as Figure 1 As shown, embodiment 1 provides a method for noise classification detection of meteorological satellite images based on a visual large model, including:

[0047] Step S100, obtaining a meteorological satellite image for preprocessing, and cropping it according to a preset method to generate a plurality of meteorological satellite image blocks;

[0048] Specifically, each meteorological satellite image contains a large-scale imaging area, usually covering hundreds of square kilometers. In order to enable the model to process these large-scale images, it is necessary to crop the large-scale meteorological satellite images to generate meteorological satellite image blocks. First, it is necessary to generate multiple meteorological satellite image blocks for the large-scale meteorological satellite images so that the model can process these images.

[0049] Based on the above analysis, it can be seen that the large-scale meteorological satellite image is effectively cropped into multiple small-scale meteorological satellite image blocks, which provides a basis for subsequent noise recognition, detection and other processing. This preprocessing method not only improves the computational efficiency, but also helps to capture local detail features in the image.

[0050] Step S200, performing noise type identification on each of the meteorological satellite image blocks based on a preset identification model to generate an identification result;

[0051] Specifically, a pre-trained noise type recognition model needs to be loaded. Next, multiple pre-processed meteorological satellite image blocks are input into the recognition model. Each image block will be processed individually by the model to determine its noise type. The model will extract and classify the input image blocks, and determine whether the image blocks contain noise points, noise lines, or other types of noise based on the learned noise features. Finally, the model will output the noise type recognition result for each image block. This result is usually expressed in the form of a label, such as "noise points", "noise lines", or "no noise".

[0052] Based on the above analysis, it can be seen that compared with manual inspection, the noise type recognition method based on the preset recognition model can greatly improve the processing efficiency and quickly complete the noise type judgment of a large number of image blocks. Manual inspection of noise types is easily affected by factors such as fatigue and subjective judgment, resulting in a high error rate. The method based on the preset recognition model can maintain consistent judgment standards and reduce human errors. Once the noise type recognition model is trained, it can be easily applied to other similar meteorological satellite images and has strong scalability. In addition, as new data is continuously added, the model can also be continuously optimized and updated to maintain its accuracy and adaptability. By generating accurate noise type recognition results, the image blocks containing noise can be further processed and analyzed as needed, such as denoising and repair. This helps to improve the quality and availability of meteorological satellite images and provide more valuable information for subsequent applications such as weather forecasting and disaster prediction.

[0053] Step S300, based on the recognition result, using a preset detection model to perform detection and generate a noise detection result;

[0054] Specifically, it is necessary to load preset detection models corresponding to the noise type. These models are specially trained for different types of noise such as noise points and noise lines, and can locate and quantify noise more accurately. According to the noise type recognition results, the image blocks containing noise points or noise lines are screened out. This can avoid unnecessary processing of noise-free image blocks and improve processing efficiency. The screened noisy image blocks are input into the corresponding detection models. These models will further analyze the details of the image blocks, locate and quantify the noise areas therein. The detection model will output detailed noise detection results for each noisy image block. These results are usually expressed in the form of coordinates, size, intensity, etc. of the noise area, providing important information for subsequent image restoration or analysis.

[0055] Based on the above analysis, it can be seen that the accuracy of noise detection can be significantly improved by processing noisy image blocks with dedicated detection models. These models are trained with a large amount of labeled data and can learn the subtle features of noise and accurately locate and quantify it. The generated noise detection results contain detailed information about the noise area, such as coordinates, size, intensity, etc. This information is of great value for subsequent image restoration, analysis or application, and can help users understand the distribution and impact of noise more deeply. The noise detection method based on the recognition results can avoid unnecessary processing of noise-free image blocks, thereby optimizing the entire processing flow. This not only saves computing resources, but also improves processing efficiency. The entire noise detection and recognition process is automated without manual intervention. This greatly reduces the cost and error rate of manual operation and improves the reliability and stability of the entire system.

[0056] In this embodiment, step S100 includes:

[0057] Step S101, obtaining a large-scale meteorological satellite image;

[0058] Specifically, a large-scale meteorological satellite image is obtained from a storage medium of the satellite, for example, a meteorological satellite image covering the entire European region is obtained, with a size of 10000×10000 pixels.

[0059] Based on the above analysis, this step ensures the original data source of meteorological satellite images and provides a rich information basis for subsequent processing. Large-scale images usually contain wider geographical areas and detailed information, which helps to analyze meteorological phenomena more comprehensively.

[0060] Step S102, performing preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image;

[0061] Specifically, preprocessing may include, for example, image denoising, image normalization, etc. Before cropping, the original image (large-scale meteorological satellite image) is subjected to preliminary denoising to reduce the impact of noise on subsequent analysis, for example, by applying various image filtering algorithms (such as Gaussian filtering, median filtering, etc.). After denoising, the pixel values ​​of the image are normalized to a certain range (such as 0-1) for subsequent processing and analysis. Normalization helps to eliminate brightness differences between different images and improve the stability of the algorithm.

[0062] Based on the above analysis, it can be seen that the preprocessing step helps to eliminate or reduce noise, distortion or other interference factors in the original image and improve the image quality, which provides more accurate input data for subsequent analysis and detection and enhances the reliability of the results.

[0063] Step S103, based on the processed large-scale meteorological satellite image, cropping is performed according to a preset method to generate the plurality of meteorological satellite image blocks;

[0064] Specifically, before cropping, for example, the cropping strategy can be determined first, including parameters such as cropping size and step size. The cropping size determines the size of each image block, and the step size determines the degree of overlap between adjacent image blocks. Subsequently, for example, a sliding window method can be used to slide on the original image, and multiple meteorological satellite image blocks can be cropped according to the determined cropping size, and the position of the sliding window is controlled by the step size.

[0065] In a specific embodiment, in order to balance computational efficiency and image details, a cropping size of 512x512 pixels and a step size of 256 pixels are selected, which means that each image block will cover an area of ​​512x512 pixels and there will be an overlap of 256 pixels between adjacent image blocks. For example, a sliding window is applied to a meteorological satellite image of the European region, and multiple image blocks are cropped according to a cropping size of 512x512 pixels and a step size of 256 pixels, which will generate a series of small-scale image blocks covering the entire European region.

[0066] Based on the above analysis, it can be seen that cropping a large-scale image into multiple small-scale image blocks makes subsequent processing more efficient and flexible. This block processing not only reduces the demand for computing resources, but also allows for more detailed analysis of different areas of the image, thereby improving the accuracy and pertinence of detection. In addition, through the preset cropping method, the overlap or connection between image blocks can be ensured to avoid information loss or disconnection.

[0067] In this embodiment, step S200 includes:

[0068] Step S201, obtaining a preset recognition model, wherein the preset recognition model includes a noise point image recognition model and a noise line image recognition model;

[0069] Specifically, step S201 is an important step in the noise type identification process, which involves loading or obtaining a pre-trained recognition model. These preset recognition models are specially designed for common noise types (such as noise points, noise lines, etc.) in meteorological satellite images, and can automatically learn and identify noise features in images. For example, the image data of the Fengyun-3 meteorological satellite is being processed, and these images may be affected by various noises. In order to identify the type of noise in the image, the recognition model needs to be trained in advance. In step S201, two pre-trained models are obtained: one is a noise image recognition model, and the other is a noise line image recognition model. Both models are built based on deep learning technology (such as convolutional neural network CNN or Transformer) and are trained using a large amount of annotated meteorological satellite image data. These models have learned how to distinguish between normal areas in the image and areas containing noise points or noise lines.

[0070] Based on the above analysis, it can be seen that step S201 provides a powerful tool for noise type identification of meteorological satellite images by obtaining a preset recognition model. These pre-trained models are learned based on a large amount of labeled data and can accurately identify different types of noise such as noise points and noise lines in the image. The beneficial effect of this method is that it significantly improves the accuracy and efficiency of noise identification. First, the preset recognition model avoids the time consumption of real-time training models, making the entire recognition process faster; secondly, because the model has been fully trained, it can more accurately identify noise features and reduce the possibility of misjudgment and missed judgment; finally, the scalability of these preset models allows continuous optimization and updating with the addition of new data to adapt to the ever-changing noise types and patterns, ensuring the continued effectiveness of the recognition system. Therefore, step S201 provides an accurate, efficient and sustainable solution for noise detection of meteorological satellite images by introducing a preset recognition model.

[0071] Step S202, performing noise type identification on each of the meteorological satellite image blocks based on the noise point image identification model and the noise line image identification model, and generating a noise point identification result, a noise line identification result and a noise-free result;

[0072] Specifically, after the preset recognition model is obtained in step S201, step S202 is to use these models to identify the noise type of the preprocessed and cropped meteorological satellite image blocks. The purpose of this step is to determine whether each image block contains noise points, noise lines or other types of noise, so as to provide a basis for subsequent targeted noise detection and removal. For example, the image processing of the Fengyun-3 meteorological satellite is continued as an example. In step S202, the cropped meteorological satellite image blocks are input one by one into the noise image recognition model and the noise line image recognition model obtained in step S201. The noise image recognition model will analyze the pixels in the image block to determine whether there are isolated points similar to salt and pepper noise; while the noise line image recognition model will detect whether there are long strips in the image that are greatly different from the surrounding pixels. These areas are usually caused by noise lines. For example, if an image block contains noise lines caused by sensor failure, the noise line image recognition model will recognize this feature and output the corresponding recognition result (such as "noise line"). Similarly, if there is noise in the image block due to electromagnetic interference, the noise image recognition model will also output a corresponding recognition result (such as "noise"). If there is no obvious noise feature in the image block, both models will output a "noise-free" result. In this way, the type of noise in each image block can be accurately identified.

[0073] Based on the above analysis, it can be seen that step S202 identifies the noise type of each meteorological satellite image block based on the pre-trained noise image recognition model and the noise line image recognition model, and this process brings significant beneficial effects. First, by using these highly specialized models, the system can accurately identify whether the image block contains noise points, noise lines or other types of noise, providing a reliable basis for subsequent targeted noise processing and image restoration. Secondly, this recognition process has a high degree of automation, which greatly improves processing efficiency and accuracy, reduces the need for manual intervention, and reduces the possibility of human error. Finally, based on the recognition results of the noise type, the system can more intelligently select the corresponding noise detection and processing methods, thereby further improving the effect of the entire noise detection and removal process. Therefore, step S202 provides important support for improving the quality and availability of meteorological satellite images through accurate identification of noise types.

[0074] In this embodiment, step S300 includes:

[0075] Step S301, obtaining a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model;

[0076] Specifically, according to the noise type (such as noise points or noise lines) identified in step S202, the corresponding preset detection model is obtained. For example, after step S202, it is identified that a certain meteorological satellite image contains both noise points and noise lines. In step S301, the system will load the preset detection models for noise points and noise lines respectively. These models are trained on a meteorological satellite image data set containing various types of noise points and noise lines, so they can accurately identify and quantify noise points and noise lines in the image.

[0077] Based on the above analysis, it can be seen that step S301 provides a targeted solution for noise detection of meteorological satellite images by obtaining a preset detection model. This step accurately selects the corresponding detection model based on the results of the previous noise type identification, thereby ensuring the accurate detection and quantification of specific types of noise (such as noise points or noise lines). This targeted detection method greatly improves the accuracy and efficiency of noise detection, and provides a reliable foundation for subsequent image quality assessment, denoising or data analysis. In addition, the scalability and updatability of the preset detection model enable the system to adapt to the ever-changing noise types and patterns, and maintain the continued effectiveness of the detection system. Therefore, step S301 provides an efficient, accurate and sustainable solution for noise detection of meteorological satellite images by accurately selecting and applying the preset detection model.

[0078] Step S302, based on the noise recognition result, use the image noise detection model to perform detection to generate a noise detection result;

[0079] Specifically, in step S302, the system inputs the image blocks containing noise into the image noise detection model according to the noise recognition results output by the noise image recognition model in step S202. The image noise detection model will further analyze and process these image blocks to generate detailed noise detection results. For example, if step S202 identifies that a certain meteorological satellite image contains noise, then in step S302, the system will input the image into the image noise detection model. The image noise detection model will analyze the noise features in the image, such as the size, density, and distribution of the noise, and generate corresponding noise detection results. The noise detection results may, for example, include information such as the number, location, and intensity of the noise, providing a detailed basis for subsequent denoising processing.

[0080] Based on the above analysis, it can be seen that step S302 realizes the precise positioning and quantification of noise points in meteorological satellite images by using the image noise detection model to detect based on the noise recognition results. This step ensures that only image blocks containing noise points are further processed, thereby avoiding unnecessary waste of computing resources and improving processing efficiency. At the same time, the image noise detection model can deeply analyze and extract the characteristics of noise points, generate detailed noise detection results, and provide accurate and reliable guidance for subsequent image denoising processing. Therefore, step S302 not only improves the accuracy of noise detection, but also optimizes the processing flow, providing effective support for improving the quality of meteorological satellite images.

[0081] Step S303, based on the noise line recognition result, using the image noise line detection model to perform detection to generate a noise line detection result;

[0082] Specifically, similar to step S302, step S303 is to detect the image block containing noise lines using the image noise line detection model based on the noise line recognition result. The image noise line detection model can accurately locate and quantify the noise line area in the image and generate detailed detection results. For example, if step S202 identifies that a meteorological satellite image contains noise lines, then in step S303, the system will input the image into the image noise line detection model. The image noise line detection model will analyze the noise line features in the image, such as the length, width, direction, etc. of the noise lines, and generate corresponding noise line detection results. The noise line detection results may, for example, include information such as the number, position, shape, etc. of the noise lines, providing guidance for subsequent denoising processing.

[0083] Based on the above analysis, it can be seen that step S303 significantly improves the accuracy and detail of noise line identification in meteorological satellite images by using the image noise line detection model for detection based on the noise line recognition results. This step ensures that in-depth noise line feature analysis is only performed on image blocks containing noise lines, thereby effectively avoiding waste of resources and improving overall processing efficiency. The image noise line detection model can accurately capture key information such as the position, length, and width of the noise line, providing strong data support for subsequent image restoration or denoising processing. Therefore, step S303 not only enhances the accuracy of noise line detection, but also provides substantial help in improving the quality of meteorological satellite images.

[0084] Step S304, generating the noise detection result based on the noise point detection result and the noise line detection result;

[0085] Specifically, step S304 is an integration step, which combines the noise point detection results and noise line detection results generated in step S302 and step S303 to form a final noise detection result. The noise detection result provides comprehensive and accurate information for subsequent image quality assessment, denoising or data analysis. For example, after step S302 and step S303, the noise point detection result and noise line detection result of a certain meteorological satellite image are obtained. In step S304, the system will merge these two results into a comprehensive result (noise detection result). The noise detection result may, for example, include detailed information such as the number, position, size, and intensity of noise points and noise lines, providing strong support for subsequent image quality improvement and data analysis.

[0086] Based on the above analysis, it can be seen that step S304 generates a comprehensive noise detection result by combining the noise point detection results and the noise line detection results. This step significantly improves the integrity and accuracy of noise detection in meteorological satellite images. By merging the two different types of noise detection results, the user is provided with a detailed overview of all noise in the image, including the number, location, intensity of noise points, and the length, width and direction of noise lines. It not only facilitates the user's comprehensive understanding of image noise, but also provides an important basis for subsequent image quality assessment, denoising processing or further data analysis. Therefore, by integrating the noise point detection results and the noise line detection results, a comprehensive, accurate and practical solution is provided for noise detection in meteorological satellite images.

[0087] In practical applications, meteorological satellite images are first annotated and the data is preprocessed before being divided into training and test sets. The training set is then fed into the constructed deep neural network for training. After the model training is completed (noise point image recognition model and noise line image recognition model are generated), when noise detection is performed on each meteorological satellite image, the overall process is as follows: Figure 4 As shown. First, the input meteorological satellite image is cropped into multiple different meteorological satellite image blocks. Then, the noise image recognition model and the noise line image recognition module are used for each image block to determine whether the image contains noise or noise line data. Next, the image noise detection module and the image noise line detection module are used for the noise image and the noise line image respectively to identify the noise data of each image (i.e., generate noise detection results).

[0088] Specifically, the meteorological satellite image noise classification detection method based on the visual large model of the present invention includes:

[0089] Step 1: Data preprocessing to generate multiple meteorological satellite image blocks;

[0090] Each meteorological satellite image contains a large-scale imaging area, usually covering hundreds of square kilometers. In order to enable the model to process these large-scale images, it is necessary to crop the large-scale meteorological satellite images to generate multiple meteorological satellite image blocks. First, it is necessary to generate small meteorological satellite image blocks for the large-scale meteorological satellite images so that the model can process these images. These cropped meteorological satellite image blocks provide a data basis for the subsequent use of noise point and noise line recognition and noise point and noise line detection.

[0091] Step 2: Build a noise point image recognition model and a noise line image recognition model based on the Transformer network;

[0092] The present invention relates to a noise image recognition model and a noise line image recognition model. The noise image recognition model and the noise line image recognition model adopt the same network architecture, and use different data to train the corresponding models respectively. The network architecture for noise or noise line recognition is mainly composed of Transformer network modules. The Transformer network architecture utilizes the self-attention mechanism, which makes it easier to capture global dependencies and establish local dependencies. The Transformer architecture used in the present invention is DeiT. Compared with the original architecture, DeiT shows higher efficiency and stronger robustness, and is a lightweight Transformer network architecture. The network obtains the same recognition ability from a large visual model through a distillation strategy, but has a smaller model size. Finally, the model uses category embedding and distilled embedding to jointly recognize noise and noise line images.

[0093] Specifically, the DeiT network first divides each input image into non-overlapping blocks, and uses a linear mapping layer to represent each block as a block token. Secondly, category tokens and distillation tokens are added at the front and back of each block token to represent the category semantic information and distillation semantic information of the image, respectively. Then, all token features are repeatedly used for several times of self-attention layers and feedforward network layers to learn and represent the semantic information they contain. Then, the final category embedding representation and distillation embedding representation of the input image are obtained. Finally, the cross entropy loss function and the Kullback–Leibler divergence loss function are used to learn the image category semantic information and the semantic information contained in the large model network, respectively. Their combination is the final loss of the training model. The network used in the present invention obtains the same recognition ability from the large visual model through a distillation strategy, but has a smaller model size. Finally, the model can use category embedding and distillation embedding to effectively recognize noisy and line images together.

[0094] Step 3: Image noise point detection model and image noise line detection model based on the visual big model;

[0095] The image noise detection model and image noise line detection model based on the visual big model use the same network structure and are trained using different data. The noise and noise line detection model uses the designed network structure to predict the noise and noise line areas by inputting the corresponding point prompt words and satellite images. The designed network structure consists of an image encoder based on the visual big model network, a prompt word encoder based on the convolutional network, and a noise and noise line Mask image decoding module, such as Figure 5 shown.

[0096] The image encoder based on the visual large model network uses a pre-trained Transformer network with strong scalability and representation capabilities, such as the SAM image encoding network pre-trained by Meta. The pre-trained network first divides each image into several blocks, and uses a convolutional network to extract the feature map of each block. Secondly, a position embedding representation is added after the feature map. Then, these feature maps are processed using several self-attention layers and feedforward network layers. Finally, several layers of convolutional layers are used to obtain the final representation of the input image. Through the self-attention module, the encoding network can effectively process long-distance dependencies and capture global information.

[0097] The prompt word encoder based on convolutional network mainly encodes the prompt words in dot format. The prompt word encoder uses several layers of convolutional neural network for encoding, which can effectively process data with local correlation, and realize efficient extraction and encoding of prompt word image features through weight sharing and pooling operations. The image embedding and prompt word embedding are obtained respectively through the image encoder and the prompt word encoder. The two embedded representations are fused, for example, by splicing, adding and other operations. Then the noise point and noise line Mask image decoding module is used to decode the fused features to generate the noise point and noise line Mask result.

[0098] The noise point and noise line Mask image decoding module also adopts the Transformer network structure, which can effectively predict whether there is noise at each pixel position. The Transformer is implemented using several self-attention layers, and finally uses a convolutional layer and a feedforward neural network to output the predicted mask image and the predicted probability value. In short, the present invention first determines the noise type of the image based on the noise point or noise line recognition model of the Transformer network. Then, the corresponding noise detection model is used to more efficiently detect the corresponding noise data, effectively improving the noise detection accuracy.

[0099] In summary, the meteorological satellite image noise classification detection method based on the visual big model of the present invention uses the Transformer network to identify the classification categories of noise on the meteorological satellite image, namely noise points and noise lines. On the basis of the classification categories, the image noise point detection model and the image noise line detection model based on the visual big model are used to specifically detect the area of ​​noise information. By combining the two, the accuracy of noise detection in meteorological satellite images is effectively improved in a hierarchical identification and detection manner. The image noise point detection model and the image noise line detection model based on the visual big model use point prompt word information and combine the visual big model to detect the noise area. On the basis of noise point and noise line image recognition, the present invention effectively improves the generalization ability and detection accuracy of the meteorological satellite image noise detection model by utilizing the prompt information and the generalization ability of the visual big model.

[0100] Embodiment 2

[0101] Figure 2 is a schematic diagram of the structure of a meteorological satellite image noise classification detection system based on a visual large model in Embodiment 2 of the present invention, such as Figure 2 As shown, Embodiment 2 provides a meteorological satellite image noise classification detection system based on a visual large model, including: a first generation module 201, a second generation module 202, and a third generation module 203. The first generation module 201 is used to obtain a meteorological satellite image for preprocessing, and to crop it according to a preset method to generate a plurality of meteorological satellite image blocks. The second generation module 202 is used to identify the noise type of each meteorological satellite image block based on a preset recognition model to generate a recognition result. The third generation module 203 is used to perform detection based on the recognition result using a preset detection model to generate a noise detection result.

[0102] In this embodiment, the first generation module 201 includes: a first acquisition unit, a first generation unit, and a second generation unit. The first acquisition unit is used to acquire a large-scale meteorological satellite image. The first generation unit is used to perform preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image. The second generation unit is used to perform cropping based on the processed large-scale meteorological satellite image according to a preset method to generate the multiple meteorological satellite image blocks.

[0103] In this embodiment, the second generation module 202 includes: a second acquisition unit and a third generation unit. The second acquisition unit is used to acquire a preset recognition model, wherein the preset recognition model includes a noise image recognition model and a noise line image recognition model. The third generation unit is used to identify the noise type of each meteorological satellite image block based on the noise image recognition model and the noise line image recognition model, and generate a noise recognition result, a noise line recognition result and a noise-free result.

[0104] In this embodiment, the third generation module 203 includes: a third acquisition unit, a fourth generation unit, a fifth generation unit and a sixth generation unit. The third acquisition unit is used to acquire a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model. The fourth generation unit is used to perform detection based on the noise point recognition result using the image noise point detection model to generate a noise point detection result. The fifth generation unit is used to perform detection based on the noise line recognition result using the image noise line detection model to generate a noise line detection result. The sixth generation unit is used to generate the noise detection result based on the noise point detection result and the noise line detection result.

[0105] The various variations and specific examples of the meteorological satellite image noise grading detection method based on a visual large model provided in Example 1 are also applicable to the meteorological satellite image noise grading detection system based on a visual large model provided in this embodiment. Through the above detailed description of a meteorological satellite image noise grading detection method based on a visual large model, those skilled in the art can clearly know the implementation method of a meteorological satellite image noise grading detection system based on a visual large model in this embodiment, so for the sake of brevity of the specification, it will not be described in detail here.

[0106] Embodiment 3

[0107] Figure 3 is a schematic diagram of the structure of an electronic device in Embodiment 3 of the present invention, such as Figure 3 As shown, the third embodiment further provides an electronic device 300 , which may include: a processor 301 and a memory 302 .

[0108] The memory 302 is used to store programs; the memory 302 may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory 302 is used to store computer programs (such as applications, functional modules, etc. that implement the above method), computer instructions, etc., and the above computer programs, computer instructions, etc. can be partitioned and stored in one or more memories 302. And the above computer programs, computer instructions, data, etc. can be called by the processor 301.

[0109] The above-mentioned computer programs, computer instructions, etc. may be stored in partitions in one or more memories 302 . And the above-mentioned computer programs, computer instructions, etc. may be called by the processor 301 .

[0110] The processor 301 is used to execute the computer program stored in the memory 302 to implement each step of the method involved in the above embodiment.

[0111] For details, please refer to the relevant description in the previous method embodiment.

[0112] The processor 301 and the memory 302 may be independent structures or integrated structures. When the processor 301 and the memory 302 are independent structures, the memory 302 and the processor 301 may be coupled and connected via a bus 303 .

[0113] The electronic device of this embodiment can execute the technical solution in the above method, and its specific implementation process and technical principle are the same, which will not be repeated here.

[0114] Embodiment 4

[0115] Embodiment 4 also provides a computer-readable storage medium, including a computer program and instructions. When the computer program or instructions are executed on a computer, the computer executes the method for noise classification detection of meteorological satellite images based on a visual large model according to any embodiment of the present invention.

[0116] Computer-readable storage media include: USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks, and other media that can store program codes.

[0117] This embodiment also provides a computer program product, which includes: a computer program, which is stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and at least one processor executes the computer program so that the electronic device executes the solution provided by any of the above embodiments.

[0118] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0119] In summary, the meteorological satellite image noise hierarchical detection method and system based on the visual big model of the present invention can accurately identify whether the meteorological satellite image contains noise points or noise lines through the noise point or noise line image recognition model based on the Transformer network, thereby providing a basis for subsequent noise detection; using the meteorological satellite image noise point or image noise line detection model based on the visual big model, the deep semantic features of noise can be effectively learned to improve the accuracy of noise detection; through the hierarchical recognition and detection method, the noise type is first identified and then the corresponding detection model is applied, thereby improving the efficiency and accuracy of noise detection.

[0120] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for noise classification detection of meteorological satellite images based on a visual large model, characterized in that: include: Step S100, obtaining a meteorological satellite image for preprocessing, and cropping it according to a preset method to generate a plurality of meteorological satellite image blocks; Step S200, performing noise type identification on each of the meteorological satellite image blocks based on a preset identification model to generate an identification result; Step S300, based on the recognition result, using a preset detection model to perform detection and generate a noise detection result; The preset recognition model adopts the Transformer-based DeiT network structure; Wherein, the step S300 includes: Step S301, obtaining a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model, and the image noise point detection model and the image noise line detection model use the same network structure and are trained using different data; Step S302, based on the noise recognition result, use the image noise detection model to perform detection to generate a noise detection result; Step S303, based on the noise line recognition result, using the image noise line detection model to perform detection to generate a noise line detection result; Step S304, generating the noise detection result based on the noise point detection result and the noise line detection result; The image noise point detection model and the image noise line detection model predict the noise point area and the noise line area by inputting the corresponding point prompt words and the meteorological satellite image block using the network structure; The network structure is composed of an image encoder based on a large visual model network, a prompt word encoder based on a convolutional network, and a noise point and line Mask image decoding module.

2. The method for detecting noise levels in meteorological satellite images based on a visual large model according to claim 1, characterized in that: The step S100 includes: Step S101, obtaining a large-scale meteorological satellite image; Step S102, performing preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image; Step S103 , based on the processed large-scale meteorological satellite image, cropping is performed according to a preset method to generate the plurality of meteorological satellite image blocks.

3. The method for detecting noise levels of meteorological satellite images based on a visual large model according to claim 1, characterized in that: The step S200 includes: Step S201, obtaining a preset recognition model, wherein the preset recognition model includes a noise point image recognition model and a noise line image recognition model; Step S202: Based on the noise point image recognition model and the noise line image recognition model, noise type recognition is performed on each meteorological satellite image block to generate a noise point recognition result, a noise line recognition result and a noise-free result.

4. A meteorological satellite image noise classification detection system based on a visual large model, characterized in that: include: The first generation module is used to obtain a meteorological satellite image for preprocessing and crop it according to a preset method to generate a plurality of meteorological satellite image blocks; A second generating module is used to identify the noise type of each meteorological satellite image block based on a preset recognition model to generate a recognition result; A third generating module is used to perform detection based on the recognition result using a preset detection model to generate a noise detection result; The preset recognition model adopts the Transformer-based DeiT network structure; Wherein, the third generation module includes: A third acquisition unit is used to acquire a preset detection model, wherein the preset detection model includes an image noise point detection model and an image noise line detection model based on a large visual model, and the image noise point detection model and the image noise line detection model adopt the same network structure and are trained using different data; A fourth generating unit, configured to generate a noise detection result by performing detection using the image noise detection model based on the noise recognition result; A fifth generating unit, configured to generate a noise line detection result by using the image noise line detection model to perform detection based on the noise line recognition result; a sixth generating unit, configured to generate the noise detection result based on the noise point detection result and the noise line detection result; The image noise point detection model and the image noise line detection model predict the noise point area and the noise line area by inputting the corresponding point prompt words and the meteorological satellite image block using the network structure; The network structure is composed of an image encoder based on a large visual model network, a prompt word encoder based on a convolutional network, and a noise point and line Mask image decoding module.

5. The meteorological satellite image noise classification detection system based on visual large model as claimed in claim 4 is characterized in that: The first generation module comprises: A first acquisition unit is used to acquire large-scale meteorological satellite images; A first generating unit is used to perform preprocessing based on the large-scale meteorological satellite image to generate a processed large-scale meteorological satellite image; The second generating unit is used to perform cropping based on the processed large-scale meteorological satellite image according to a preset method to generate the multiple meteorological satellite image blocks.

6. The meteorological satellite image noise classification detection system based on visual large model as claimed in claim 4 is characterized in that: The second generation module comprises: A second acquisition unit, used to acquire a preset recognition model, wherein the preset recognition model includes a noise point image recognition model and a noise line image recognition model; The third generating unit is used to identify the noise type of each meteorological satellite image block based on the noise point image recognition model and the noise line image recognition model, and generate a noise point recognition result, a noise line recognition result and a noise-free result.

7. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively coupled to the at least one processor; Wherein, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the meteorological satellite image noise classification detection method based on the visual large model as described in any one of claims 1-3.

8. A computer-readable storage medium, characterized in that: It comprises a computer program and instructions. When the computer program or the instructions are run on a computer, the computer is enabled to execute the noise classification detection method for meteorological satellite images based on a visual large model as described in any one of claims 1 to 3.

Images